Fig. 1. Overview of DLM. In Arm Curriculumn in Circle 1 (blue area) we formulated a randomization course for robotic arms. This part of the robotic arms has no collision volume in the simulation. The grasping postures of different robotic arms are generated through various randomization methods. The difficulty gradually transitions from the simplest fixed starting posture to the most difficult random linear interpolation posture; in Arm Estimator in Circle 2 (green area) Arm Estimator, the above Arm Curriculumn is used to achieve different actions, and the two indicators of the end effector position of the robotic arm and the center of mass position of the robotic arm are used to better assist the lower limb balance; Circle 3 (grey area) is the Sim Train process of the overall solution. Historical observation information outputs the estimated value through historical MLP, and the robotic arm information in the privileged observation information (including the center of mass of the robotic arm and ee. Pos) and ontology information (base height and linear speed) are supervised and learned, and the motor parameters are restricted through the NP3O network, and the robot's balance is finally controlled; Circle 4 (white area) is the deployment process, and the upper limbs control the movement of the robotic arm through Diffusion Policy or remote operation, while the lower limbs use the above-mentioned training network for smooth control. During the deployment process, detailed calibrations such as motor properties, sensor delay and motor torque compensation are also required to reduce sim2real's gap.